Data annotation method and device based on artificial intelligence large model

Abstract

The invention provides a data annotation method and device based on an artificial intelligence large model. The method comprises the steps of obtaining to-be-annotated data and cue words; according to the method, data to be labeled are preprocessed, original and disordered multi-modal data are converted into clean and normative data suitable for model digestion and learning, and the labeling difficulty of subsequent models is reduced. On the basis of the cue word, the preprocessed to-be-labeled data is input into a data labeling model, a pre-labeling result, output by the data labeling model, of the to-be-labeled data is obtained, and the data labeling model is obtained through training on the basis of the sample data, the manual labeling result corresponding to the sample data and the cue word corresponding to the sample data. According to the method, the data is automatically labeled through the data labeling model, so that the dependence on professionals is reduced, the labor cost is reduced, the labeling efficiency is improved, and in addition, the data leakage risk existing in a data labeling outsourcing mode is also avoided.

Description

Technical Field

[0001] This invention relates to the field of data annotation technology, and in particular to a data annotation method and apparatus based on a large artificial intelligence model. Background Technology

[0002] Against the backdrop of rapid development in large-scale modeling technology, the data annotation industry is undergoing a profound transformation from labor-intensive to knowledge-intensive, with the core shift from quantitative expansion to qualitative leap. Currently, traditional data annotation technologies suffer from the following problems:

[0003] 1. Large models often process data that is a mixture of multiple modalities, such as text, images, audio, video, and even 3D point clouds. The annotation work is not limited to a single modality, but more importantly, it is necessary to achieve accurate alignment and semantic association across modalities, which greatly increases the technical complexity and annotation difficulty.

[0004] 2. Faced with increasingly complex data annotation tasks, there is a growing reliance on compound talents who understand both AI and vertical domain knowledge, which is costly and the annotation efficiency is limited by manpower.

[0005] 3. Training large models involves a large amount of sensitive data (such as medical records, personal voice recordings, and street view images). Traditional internet crowdsourcing models pose a significant risk of data leakage.

[0006] To address the aforementioned problems in existing technologies, a new data annotation method is urgently needed. Summary of the Invention

[0007] This invention provides a data annotation method and apparatus based on a large artificial intelligence model to solve the problems of high difficulty in data annotation, high labor costs, low manual efficiency, and risk of data leakage in the prior art.

[0008] This invention provides a data annotation method based on a large artificial intelligence model, comprising:

[0009] Obtain the data to be labeled and the prompt words;

[0010] The data to be labeled is preprocessed;

[0011] Based on the prompt words, the preprocessed data to be labeled is input into the data labeling model, and the pre-labeling result of the data to be labeled is obtained from the data labeling model.

[0012] The data annotation model is trained based on sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0013] According to the present invention, a data annotation method based on a large artificial intelligence model is provided, wherein the data annotation model is obtained through the following training method:

[0014] A training dataset is constructed based on the technical problem to be solved; the training dataset includes sample data, the manually labeled results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0015] Based on the sample data, the corresponding manual annotation results, and the corresponding prompt words, the data annotation model is trained in a supervised manner.

[0016] According to the data annotation method based on a large artificial intelligence model provided by the present invention, after obtaining the pre-annotation results of the data to be annotated output by the data annotation model, the method further includes:

[0017] Based on preset filtering conditions, a portion of the data to be labeled and the corresponding pre-labeling results are selected and sent to the secondary labeling interface to obtain the secondary labeling results returned by the secondary labeling interface;

[0018] Based on the secondary annotation results, the current version of the data annotation model is incrementally trained or fine-tuned, and the data annotation model after incremental training or fine-tuning is used as the current version of the data annotation model.

[0019] According to the data annotation method based on a large artificial intelligence model provided by the present invention, before inputting the preprocessed data to be annotated into the data annotation model, the method further includes:

[0020] Obtain the annotation performance of the data annotation model in each historical version, and select the data annotation model with the best annotation performance as the data annotation model in the current version.

[0021] According to a data annotation method based on a large artificial intelligence model provided by the present invention, the data to be annotated is preprocessed, including:

[0022] The text data in the data to be labeled is converted into a vector rich in semantic information;

[0023] The format of the image data in the data to be labeled is standardized.

[0024] According to the present invention, a data annotation method based on a large artificial intelligence model is provided to transform the text data in the data to be annotated into a vector rich in semantic information, including: using a pipeline of regular expression cleaning - word segmentation - part-of-speech / syntactic analysis - word embedding to transform the text data in the data to be annotated into a vector rich in semantic information;

[0025] Unifying the format of image data in the data to be labeled includes:

[0026] A pipeline of Resize / Crop – Data Augmentation – Color Conversion – Normalization is used to transform the image data in the data to be labeled into a unified format.

[0027] According to the present invention, a data annotation method based on a large artificial intelligence model is provided, wherein the pre-annotation results include text annotation results and / or image annotation results;

[0028] The text annotation results are obtained through the language large model part of the data annotation model; the image annotation results are obtained through the visual large model annotation of the data annotation model.

[0029] The present invention also provides a data annotation device based on a large artificial intelligence model, comprising:

[0030] The data acquisition module is used to acquire the data to be labeled and prompt words;

[0031] The preprocessing module is used to preprocess the data to be labeled;

[0032] The data annotation module is used to input the preprocessed data to be annotated into the data annotation model based on the prompt words, and obtain the pre-annotation result of the data to be annotated output by the data annotation model;

[0033] The data annotation model is trained based on sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0034] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the data annotation method based on a large artificial intelligence model as described above.

[0035] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the data annotation method based on a large artificial intelligence model as described above.

[0036] This invention provides a data annotation method and apparatus based on a large-scale artificial intelligence model. It acquires the data to be annotated and prompt words; preprocesses the data to be annotated, transforming the raw, messy multimodal data into clean, standardized data suitable for model digestion and learning, thus reducing the difficulty of subsequent model annotation. Based on the prompt words, the preprocessed data to be annotated is input into the data annotation model, obtaining the pre-annotation results output by the model. The data annotation model is trained based on sample data, the corresponding manual annotation results, and the prompt words. This automated data annotation reduces reliance on specialized personnel, lowers labor costs, improves annotation efficiency, and avoids the data leakage risks associated with outsourced data annotation. Attached Figure Description

[0037] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0038] Figure 1 This is a flowchart illustrating the data annotation method based on a large artificial intelligence model provided by the present invention;

[0039] Figure 2 This is a schematic diagram of the model optimization and iterative training process in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the data annotation device based on a large artificial intelligence model provided by the present invention;

[0041] Figure 4 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0042] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0043] The following is combined Figure 1 and Figure 2 This invention describes a data annotation method based on a large artificial intelligence model.

[0044] like Figure 1As shown, the data annotation method based on a large artificial intelligence model provided by this invention includes the following steps:

[0045] S1. Obtain the data to be labeled and prompt words.

[0046] The data to be labeled can be of various modalities, including but not limited to text, images, audio, and video, and the prompts can be natural language commands. For example, the data to be labeled could be multiple images taken by highway cameras, and the prompt could be "label all vehicles in the images".

[0047] S2. Preprocess the data to be labeled.

[0048] The preprocessing of the data to be labeled includes: converting the text data in the data to be labeled into vectors rich in semantic information; and unifying the format of the image data in the data to be labeled.

[0049] Specifically, the text data in the data to be labeled is transformed into vectors rich in semantic information. This includes: using a pipeline of regular expression cleaning—word segmentation—part-of-speech / syntactic analysis—word embedding (preferably using an AutoTokenizer + transformers embedding model) to transform the text data in the data to be labeled into structured vectors rich in semantic information and semantically computable. Text cleaning removes irrelevant characters and redundant information; word segmentation breaks down continuous text into meaningful words or sub-words; part-of-speech / syntactic analysis removes affixes from words, retaining their core stems, and assigns corresponding grammatical categories to the words in the text, such as nouns and verbs; word embedding transforms words into numerical vectors that can be used for model computation.

[0050] To standardize the format of image data in the data to be labeled, a pipeline of Resize / Crop – Data Augmentation – Color Conversion – Normalization is used to transform the image data in the data to be labeled into a unified format. Specifically, at least one of the following processing methods can be used to process the image data in the data to be labeled: adjusting the image size (unifying the image to a specified size), cropping the image (extracting the region of interest from the image), and converting the color space (such as converting RGB color mode to grayscale mode). Again, taking the above example where the data to be labeled consists of multiple images taken by highway cameras, and the prompt is "label all vehicles in the image," in the preprocessing stage, all images can be uniformly scaled to 1024x768 pixels. For the accompanying text descriptions in the images (such as the shooting time), garbled characters will be removed, and the text will be segmented and converted into word vectors.

[0051] S3. Based on the prompt words, input the preprocessed data to be labeled into the data labeling model and obtain the pre-labeling results of the data to be labeled output by the data labeling model.

[0052] The data annotation model is trained based on sample data, corresponding manual annotations, and prompts. Before using the data annotation model, it needs to be trained to meet the task requirements. In an optional embodiment of the invention, the training process of the data annotation model is as follows:

[0053] Step 1: Construct a training dataset based on the technical problem to be solved; the training dataset includes sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0054] Specifically, before constructing the training dataset, it is required to clearly define the task objectives, problem type, data format, and annotation specifications. For example, by filling in metadata such as "dataset type," "version," "annotation type," and "annotation dictionary," a foundation is laid for subsequent automated processing. In an optional embodiment of this invention, the training dataset may include, but is not limited to, image data in formats such as JPG, JPEG, and PNG, and text data in formats including but not limited to JSON, JSON1, and CSV, possessing compatibility with multimodal data. It supports viewing imported data, including file names and annotation status. This step is the "blueprint" stage of the entire annotation process, ensuring that all subsequent work is consistent and standardized. It achieves standardized data entry and preliminary management. Large models have a huge number of parameters, and even small annotation deviations can be amplified exponentially during training, causing the model output to deviate significantly from expectations. Defining clear annotation specifications (annotation dictionary) from the source is a prerequisite for ensuring the consistency and accuracy of the final data, avoiding large-scale rework due to misunderstandings. In addition, a unified data access and management interface provides a foundation for processing highly customized, multi-source fragmented data. The data in the training dataset also needs to undergo the preprocessing described in step S2 above, transforming text data into vectors rich in semantic information and standardizing the format of image data. Furthermore, data augmentation (e.g., through rotation, flipping, scaling, etc.) can be used to expand the training dataset, indirectly improving the quality and generalization ability of the data-annotated model trained in the future.

[0055] Suppose we want to create a dataset for a model used in "autonomous driving street scene understanding". In this step, the user will create a dataset named "Urban Road Multi-Object Detection-V3", type "Image Object Detection", with the annotation dictionary defined as ["Car", "Truck", "Pedestrian", "Traffic Light - Red", "Traffic Light - Green", "Cones"]. Then, tens of thousands of street images are imported in batches. This clear setup provides crucial guidance for subsequent model pre-annotation.

[0056] Step 2: Based on the sample data, the corresponding manual annotation results, and the prompt words, conduct supervised training on the data annotation model.

[0057] Taking the dataset named "Urban Road Multi-Object Detection-V3" as an example, it can be randomly divided into training, validation, and test sets in an 8:1:1 ratio. The training set is used to calculate the loss and update the model weights; the validation set is used after each epoch to evaluate the model's performance on unseen data, monitoring the training process, preventing overfitting, and selecting the best model; the test set (data that has not been used at all during training and validation) is used for fair cross-validation of all trained model versions. During training, the model's loss value, evaluation metrics, etc., are recorded, and the model is saved after training.

[0058] The pre-annotation results of the data annotation model include text annotation results and / or image annotation results. Text annotation results are obtained through the language large-scale model component of the data annotation model; image annotation results are obtained through the visual large-scale model component. Traditional annotation typically only requires simple classification or bounding box annotation of images and text (e.g., "This is a cat," "This is a red light"), a mechanical and repetitive task. However, the large-scale model requires annotation of higher-order semantic content such as "chain-of-thought," reasoning logic, and human preferences. Chain-of-thought is a capability or technological paradigm of the large language model in solving complex reasoning tasks, explicitly generating intermediate reasoning steps. Its core idea is to mimic the human process of "step-by-step thinking," breaking down complex problems into multiple logically coherent sub-steps, and then deriving the final answer based on these intermediate steps. Therefore, in an optional embodiment of the present invention, the language model in the data annotation model uses Transformer as its basic network structure. It can combine and integrate data from different modalities such as text, images, and audio to achieve effective generation and conversion of data between different modalities. This allows for the generation of coherent, human-like text language, thus satisfying diverse text processing tasks. It possesses core algorithmic capabilities such as language understanding, information extraction, text generation, context understanding, multi-turn dialogue, and logical reasoning, and can be used to complete "thought chains" to handle highly complex annotation tasks. The visual model in the data annotation model is based on the Vision Transformer architecture and deep learning technology for computer vision algorithms, and can handle visual processing tasks such as visual detection, visual question answering, and character recognition.

[0059] The data annotation model trained through "visual-language alignment" has acquired the ability to process multimodal data and understand human natural language instructions (prompts). It can understand complex instructions such as "describe the sequence of actions of people in an image and infer their intentions," achieving cross-modal alignment annotation. Furthermore, it can be supplemented by traditional algorithm models such as OCR, Canny, SIFT, watershed segmentation, threshold segmentation, linear regression, SVM, and clustering. These algorithms are efficient and low-cost for specific tasks (such as text extraction and simple image segmentation), complementing the large model and enabling rapid customization and recognition of labeled data. Based on user prompts, it can intelligently extract text content from images, forming a crucial step in AI-automated pre-annotation.

[0060] The data annotation model, trained with "visual-language alignment," can understand and annotate "reasoning logic" and "thought chains." For example, given a large medical text, it can automatically annotate the reasoning chain of "symptoms -> preliminary diagnosis -> recommended examinations." Furthermore, it can handle knowledge-based tasks requiring specialized backgrounds, reducing reliance on a large number of expert annotators. Ordinary annotators only need to review the AI ​​results, rather than annotating from scratch, significantly reducing labor costs and improving annotation efficiency. Through platform automation, manual data manipulation is reduced, requiring only minimal manual review and fine-tuning of annotation effects. This simplifies the manual processes of importing, exporting, and transmitting datasets over the internet, ensuring data security and compliance.

[0061] Taking a medical data annotation task as an example, the task is to annotate a set of "patient chief complaints" text with a "diagnostic logic chain". The data annotation model can receive prompts such as: "Please analyze the following patient chief complaints and generate a thought chain containing 'symptom extraction', 'possible etiology inference', and 'suggested examination items'." The model output might be: Symptom extraction: fever, cough, chest pain -> Possible etiology inference: community-acquired pneumonia is highly likely -> Suggested examination items: blood routine, chest X-ray. This greatly reduces the workload of professional physician annotators.

[0062] like Figure 2 As shown, in order to perform multiple pre-training iterations on the data annotation model to make its annotation more accurate, in an optional embodiment of the present invention, after obtaining the pre-annotation results of the data to be annotated output by the data annotation model, the method further includes:

[0063] Based on preset filtering conditions, select a portion of the data to be labeled and the corresponding pre-labeling results and send them to the secondary labeling interface to obtain the secondary labeling results returned by the secondary labeling interface;

[0064] Based on the secondary annotation results, the current version of the data annotation model is incrementally trained or fine-tuned, and the data annotation model after incremental training or fine-tuning is used as the current version of the data annotation model.

[0065] Specifically, after the data annotation model completes automatic annotation, some difficult-to-understand data annotations and a small amount of review work can be selected and sent to the secondary annotation interface. Annotators can obtain these tasks through this interface and perform manual annotation. In an optional embodiment of the present invention, the annotation tools for video and image data include, but are not limited to, rectangular boxes, polygonal boxes, AI annotation, left rotation, right rotation, zoom in, zoom out, resize, drag, clear canvas, visual multi-turn question answering, automatic extraction of image question content answers, tags, etc. The annotation tools for text data include, but are not limited to, multi-turn question answering, question answer ranking, and text corpus scenarios. These annotation tools allow humans to easily check, correct, and fine-tune the pre-annotation results of the data annotation model, instead of starting from scratch. By introducing human expertise and judgment, the quality of the data annotation model results is checked, and edge cases and highly complex scenarios that the data annotation model cannot handle are dealt with. At the same time, the manual correction results serve as high-quality feedback data to optimize the data annotation model. For example, when annotating a street view image, the data annotation model may mislabel a rare construction vehicle as a "truck". In this step, human annotators use platform tools to correct the label to "special vehicle" and submit the result. This correction itself "educates" the data annotation model. The human annotation or correction results are used as high-quality supervision signals, automatically fed back into the training dataset, triggering incremental training or fine-tuning of the model, and realizing a continuous optimization closed loop of "annotation-training-prediction-re-annotation".

[0066] In an optional embodiment of the present invention, before inputting the preprocessed data to be labeled into the data labeling model, the method further includes:

[0067] Obtain the annotation performance of the data annotation model in each historical version, and select the data annotation model with the best annotation performance as the data annotation model for the current version.

[0068] Specifically, the performance evaluation of various versions of the data annotation model is based on a unified evaluation metric system (such as accuracy, recall, F1-score, IoU, etc.). The performance of each version is recorded, and horizontal comparisons between different model versions assist users in making informed decisions about the optimal model. In an optional embodiment of this invention, different model architectures (such as Transformer, CNN, BiLSTM, etc.), data preprocessing strategies, and hyperparameter combinations can be selected to flexibly construct the optimal model structure and configuration suitable for the current annotation task. Furthermore, in an optional embodiment of this invention, model-related data can be recorded during model training, providing a complete visualization panel for training log tracking, loss curves, and metric trends. It also integrates debugging tools such as feature importance analysis and error sample attribution to help users deeply understand model behavior and accurately pinpoint optimization directions. Through the above approach, not only is single model training supported, but the goal is to build a self-evolving intelligent annotation model that continuously improves annotation accuracy and generalization ability through ongoing human-machine collaboration.

[0069] In summary, the data annotation method based on a large-scale artificial intelligence model provided by this invention can automatically preprocess text and visual data, such as automatically cleaning, segmenting, stemming, tagging, and embedding words; automatically resizing, augmenting, cropping, and converting visual images; and automatically annotating datasets such as image detection, visual question answering, and character recognition. The large-scale model automatically evaluates content analysis and recognition across multiple dimensions, including relevance, conciseness, factuality, and creativity. This method allows for a leap from simply labeling image content and object recognition to labeling higher-order semantic content such as thought chains, reasoning logic, and human preferences, reducing the need for annotators to possess logical reasoning, professional background, and even ethical judgment. Furthermore, it significantly reduces the cost of current data annotation quality control, achieving high-precision, high-consistency, and traceable annotation quality control, improving dataset quality, and promoting high-quality development of artificial intelligence. The platform automates processing and reduces manual data manipulation, requiring only minimal manual review and fine-tuning of annotation effects. It simplifies manual data import, export, and external network transmission of datasets, avoiding potential data security and compliance issues.

[0070] Based on the same inventive concept, the present invention also provides a data annotation device based on a large artificial intelligence model. The data annotation device based on a large artificial intelligence model provided by the present invention is described below. The data annotation device based on a large artificial intelligence model described below and the data annotation method based on a large artificial intelligence model described above can be referred to and corresponded to each other.

[0071] like Figure 3As shown, the data labeling device based on a large artificial intelligence model provided by the present invention includes a data acquisition module 31, a preprocessing module 32, and a data labeling module 33.

[0072] The data acquisition module 31 is used to acquire the data to be labeled and prompt words.

[0073] The preprocessing module 32 is used to preprocess the data to be labeled.

[0074] The data annotation module 33 is used to input the preprocessed data to be annotated into the data annotation model based on the prompt words, and obtain the pre-annotation result of the data to be annotated output by the data annotation model; wherein, the data annotation model is trained based on sample data, the manual annotation result corresponding to the sample data and the prompt words corresponding to the sample data.

[0075] In an optional embodiment of the present invention, the data annotation module 33 is further divided into a text data annotation submodule and an image data annotation submodule.

[0076] The text data annotation submodule is used to transform the text data in the data to be annotated into vectors rich in semantic information. By adopting a pipeline of regular expression cleaning, word segmentation, part-of-speech / syntactic analysis and word embedding (preferably using the AutoTokenizer + transformers embedding model for word embedding), the text data in the data to be annotated is transformed into a structured vector rich in semantic information and semantically computable.

[0077] The image data annotation submodule is used to unify the format of image data in the data to be annotated. It uses a pipeline of Resize / Crop - Data Augmentation - Color Conversion - Normalization to transform the image data in the data to be annotated into a unified format.

[0078] In an optional embodiment of the present invention, the above-mentioned data annotation device based on a large artificial intelligence model further includes an iterative optimization module, which is used to select a portion of the data to be annotated and the corresponding pre-annotation results based on preset screening conditions and send them to the secondary annotation interface to obtain the secondary annotation results returned by the secondary annotation interface; perform incremental training or fine-tuning on the current version of the data annotation model based on the secondary annotation results, and use the data annotation model after incremental training or fine-tuning as the current version of the data annotation model.

[0079] Figure 4 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 4As shown, the electronic device may include: a processor 410, a communication interface 420, a memory 430, and a communication bus 440, wherein the processor 410, the communication interface 420, and the memory 430 communicate with each other through the communication bus 440. The processor 410 can call logical instructions in the memory 430 to execute the data annotation method based on the large-scale artificial intelligence model provided by the above methods, which includes:

[0080] Obtain the data to be labeled and the prompt words;

[0081] The data to be labeled is preprocessed;

[0082] Based on the prompt words, the preprocessed data to be labeled is input into the data labeling model, and the pre-labeling result of the data to be labeled is obtained from the data labeling model.

[0083] The data annotation model is trained based on sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0084] Furthermore, the logical instructions in the aforementioned memory 430 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the data annotation method based on the artificial intelligence large model provided by the above methods, the method comprising:

[0086] Obtain the data to be labeled and the prompt words;

[0087] The data to be labeled is preprocessed;

[0088] Based on the prompt words, the preprocessed data to be labeled is input into the data labeling model, and the pre-labeling result of the data to be labeled is obtained from the data labeling model.

[0089] The data annotation model is trained based on sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0090] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the data annotation method based on the artificial intelligence large model provided by the above methods, the method comprising:

[0091] Obtain the data to be labeled and the prompt words;

[0092] The data to be labeled is preprocessed;

[0093] Based on the prompt words, the preprocessed data to be labeled is input into the data labeling model, and the pre-labeling result of the data to be labeled is obtained from the data labeling model.

[0094] The data annotation model is trained based on sample data, the manual annotation results corresponding to the sample data, and the prompt words corresponding to the sample data.

[0095] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0096] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0097] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence large model-based data labeling method, characterized in that, The method comprises the following steps: obtaining to-be-labeled data and prompt words; preprocessing the to-be-labeled data; inputting the preprocessed to-be-labeled data into a data labeling model based on the prompt words, and obtaining a pre-labeling result of the to-be-labeled data output by the data labeling model; wherein the data labeling model is trained based on sample data, manual labeling results corresponding to the sample data, and prompt words corresponding to the sample data. 2.The data labeling method based on an artificial intelligence large model according to claim 1, wherein, The data labeling model is obtained by the following training method: constructing a training data set according to a to-be-solved technical problem; the training data set comprises sample data, manual labeling results corresponding to the sample data, and prompt words corresponding to the sample data; supervised training the data labeling model based on the sample data, the manual labeling results corresponding to the sample data, and the prompt words corresponding to the sample data. 3.The data labeling method based on an artificial intelligence large model according to claim 1, wherein, After obtaining the pre-labeling result of the to-be-labeled data output by the data labeling model, the method further comprises the following steps: selecting part of the to-be-labeled data and the corresponding pre-labeling result based on a preset screening condition and sending them to a secondary labeling interface to obtain a secondary labeling result returned by the secondary labeling interface; incrementally training or fine-tuning the data labeling model based on the secondary labeling result, and taking the incrementally trained or fine-tuned data labeling model as the current version of the data labeling model. 4.The data labeling method based on the artificial intelligence large model according to claim 3, characterized in that, Before inputting the preprocessed to-be-labeled data into the data labeling model, the method further comprises the following steps: obtaining labeling performance of each historical version of the data labeling model, and selecting the data labeling model with the optimal labeling performance as the current version of the data labeling model. 5.The data labeling method based on an artificial intelligence large model according to claim 1, wherein, The preprocessing of the to-be-labeled data comprises the following steps: converting text data in the to-be-labeled data into vectors rich in semantic information; unifying the format of image data in the to-be-labeled data. 6.The data labeling method based on the artificial intelligence large model according to claim 5, characterized in that, Converting text data in the to-be-labeled data into vectors rich in semantic information comprises adopting a pipeline of regular cleaning-division-word-part-of-speech / syntax analysis-word embedding to convert the text data in the to-be-labeled data into vectors rich in semantic information. Unifying the format of image data in the to-be-labeled data comprises adopting a pipeline of Resize / Crop-data enhancement-color conversion-normalization to convert the image data in the to-be-labeled data into a unified format.

7. The data labeling method based on the artificial intelligence large model according to any one of claims 1-6, characterized in that, The pre-labeling result comprises a text labeling result and / or an image labeling result; The text labeling result is obtained by a language large model part in the data labeling model; and the image labeling result is obtained by a visual large model in the data labeling model.

8. An artificial intelligence large model-based data labeling apparatus, characterized by comprising: The method comprises the following steps: a data acquisition module for obtaining to-be-labeled data and prompt words; a preprocessing module for preprocessing the to-be-labeled data; a data labeling module for inputting the preprocessed to-be-labeled data into a data labeling model based on the prompt words, and obtaining a pre-labeling result of the to-be-labeled data output by the data labeling model; The data labeling model is trained based on sample data, artificial labeling results corresponding to the sample data, and prompt words corresponding to the sample data.

9. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor implements the data labeling method based on the artificial intelligence large model according to any one of claims 1 to 7 when executing the program. 10.A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the data labeling method based on the artificial intelligence large model according to any one of claims 1 to 7.

Images